Hacker News

Top Story

OpenAI is facing intense backlash after NYU mathematician Tristan Buckmaster accused the lab of using $22.5M in compute (300 billion output tokens) to brute-force a proof for the Millennium Prize Navier-Stokes problem using unreleased research techniques his team ran through Codex. The controversy has reignited fierce community debates around corporate academic poaching, LLM telemetry privacy, and whether massive compute clusters are replacing genuine mathematical insight.

Front Page Highlights

How Well Do Agents Use Test/Verification Techniques? · Dan Luu Dan Luu ran a massive empirical benchmark evaluating 26 prompt conditions and skills (including Lean 4, Verus, TDD, and property-based testing) across 80+ runs per condition on an agentic Zstd implementation. He discovered that coding agents write remarkably poor tests by default, routinely generating vacuous formal proofs (like proving A => A) or writing superficial test suites that miss critical edge cases. Consequently, default prompts with no extra testing instructions actually outperformed complex testing frameworks and skills.

We Have a Year to Fix Security Everywhere · jyn.dev With Z.ai’s open-weight GLM 5.3-flash model released and immediately “abliterated” to remove all safety refusals, frontier-class cyber-exploitation capabilities are now runnable on consumer hardware like the upcoming 256GB M5 Mac Studio. The author warns that automated exploit loops will soon scan and attack public infrastructure 24/7, leaving software maintainers and governments a narrow one-year window to mandate rapid patch deployment, memory safety, and aggressive network isolation.

Leaving VMware Just Got Harder After Broadcom Pulled VDDK Downloads · Virtualization Howto Broadcom has quietly killed public download access for VMware’s Virtual Disk Development Kit (VDDK), breaking automated migration toolchains for Nutanix Move, Red Hat OpenShift Virtualization, and Platform9. System administrators confirm that Broadcom support validated the removal as intentional, prompting widespread anger on HN over aggressive corporate lock-in tactics aimed at trapping legacy workloads on vSphere.

Replacing a Rust Enum with a 64-Bit Word Made My Interpreter 17% Faster · Pointers Gone Wild VM engineer Maxime Chevalier-Boisvert details refactoring the core value type in the Plush language interpreter from a standard Rust enum to a tagged 64-bit word utilizing float self-tagging and low-bit pointer alignment. The low-level refactoring reduced peak memory RSS by up to 37% and delivered a 17% speedup across execution benchmarks by slashing stack instruction overhead and branch dispatches.

LibreOffice Breaks Download Records After Declaring It Has No AI Features · Manual do Usuário LibreOffice 26.8 surpassed one million downloads in its first week after The Document Foundation explicitly touted “no generative AI features” as a major privacy and data-sovereignty selling point. The foundation outlined strict prerequisite principles—such as complete local execution, zero telemetry, and vendor independence—proving that an explicit “AI-free” stance is becoming a powerful market differentiator.

Jellyfin 12.0 Released · Jellyfin Blog The open-source media server dropped its legacy “10.” versioning scheme for Jellyfin 12.0, delivering long-awaited native server-side support for books and comics, an upgraded library database engine, and making the modern responsive layout default. Additional core improvements include episode version grouping, FFmpeg 8.1 integration, and targeting .NET 10.

There’s a New ‘Google Jail’ for Independent Wikis · Weird Gloop The maintainers of independent gaming wikis document a post-2024 Google Search change where brand-new root domains have all inner pages suppressed from search results, indexing only the main page for up to a year. To bypass this “Google Jail” when migrating communities off bloated corporate hosts like Fandom, wiki administrators are now forced to launch on subdomains of established root domains before redirecting later.

Show HN & Launches

OUI-1: World’s First Model for Generative UI brings a 4B-active parameter DiffusionGemma finetune capable of generating validated UI code in OpenUI Lang on consumer GPUs. By combining self-distillation with parser-based rejection rewards, the team boosted benchmark accuracy from 13.0% to 71.7% while maintaining sub-second generation speeds. On the tooling front, Show HN: DriveSync – Fast Git-Styled Google Drive Sync CLI introduces a Rust-built CLI offering git-like push, pull, and diff workflows for Google Drive, while Show HN: Copperhead – Hardware as Fast as Software pairs Claude with kicad-cli to execute automated schematic edits, netlist modifications, and ERC checks. Finally, I-Have-ADHD released a minimal skill for CLI coding agents that suppresses conversational preambles and forces action-first, numbered output.

Discussion & Debate

The day’s most intense technical debates revolved around LG TV Spying Allegations, where GamersNexus’s investigation into smart TV telemetry sparked arguments over whether LG is actively uploading offline room audio or merely storing unencrypted search logs in device memory. Despite disagreements over packet analysis methodology, the thread overwhelmingly reinforced the consensus to keep smart TVs permanently disconnected from local networks. On the infrastructure side, The VMs Powering Mobile Agents prompted deep architectural discussions comparing Claude Code’s Firecracker microVMs running isolated Rust inits against Instinct’s disposable E2B sandboxes backed by S3 Git repositories for long-term memory. Lastly, European CDN Concentration raised systemic risk alarms after data revealed that 89.6% of CDN-using European companies front their sites with Cloudflare, illustrating how single-vendor BGP or automation errors can trigger widespread regional web outages.

💡 Next Step Idea: If you’d like to explore any of these technical topics further—such as building a dedicated report on the open-source AI models mentioned or analyzing the performance benchmarking methodologies—let me know!